Data backup is a standard part of all large-scale computer data storage systems (and most small systems, as well). Data written to a primary storage medium, such as a volume on a local storage subsystem, are copied, or “mirrored,” to a backup medium, typically another volume on a remote storage subsystem. The backup volume can then be used for recovery in case a disaster causes the data on the primary medium to be lost. Methods of remote data mirroring are surveyed by Ji et al., in an article entitled “Seneca: Remote Mirroring Done Write,” Proceedings of USENIX Technical Conference (San Antonio, Tex., June, 2003), pages 253-268, which is incorporated herein by reference. The authors note that design choices for remote mirroring must attempt to satisfy the competing goals of keeping copies as closely synchronized as possible, while delaying foreground writes by host processors to the local storage subsystem as little as possible.
Large-scale storage systems, such as the IBM Enterprise Storage Server (ESS) (IBM Corporation, Armonk, N.Y.), typically offer a number of different copy service functions that can be used for remote mirroring. Among these functions is peer-to-peer remote copy (PPRC), in which a mirror copy of a source volume on a primary storage subsystem is created on a secondary storage subsystem. When an application on a host processor writes to a PPRC volume on the primary subsystem, the corresponding data updates are entered into cache memory and non-volatile storage at the primary subsystem. The control unit (CU) of the primary subsystem then sends the updates over a communication link to the secondary subsystem. When the CU of the secondary subsystem has placed the data in its own cache and non-volatile storage, it acknowledges receipt of the data. The primary subsystem then signals the application that the write operation is complete.
PPRC provides host applications with essentially complete security against single-point failures, since all data are written synchronously to non-volatile media in both the primary and secondary storage subsystems. On the other hand, the need to save all data in non-volatile storage on both subsystems before the host write operation is considered complete can introduce substantial latency into host write operations. In some large-scale storage systems, such as the above-mentioned IBM ESS, this latency is reduced by initially writing data both to cache and to high-speed, non-volatile media, such as non-volatile random access memory (RAM), in both the primary and secondary subsystems. The data are subsequently copied to disk asynchronously (an operation that is also referred to as “hardening” the data) and removed from the non-volatile memory. The large amount of non-volatile memory that must be used for this purpose is very costly.
Data mirroring functions are commonly classified as either “synchronous” or “asynchronous.” In synchronous mirroring, all updates (data write operations) are propagated immediately to the secondary subsystem. This is generally the safest mode of operation, but also the slowest, since host writes are not acknowledged until the data have been stored in non-volatile media on both the primary and secondary subsystems. When asynchronous mirroring is used, on the other hand, hosts receive notification from the primary subsystem that a write is complete as soon as the data have been placed in the cache on the primary subsystem (or possibly after the data have been secured in non-volatile storage on the primary subsystem). The updates of the data are read out from the cache and sent to the secondary subsystem in a separate, background operation. The asynchronous approach gives better performance, in terms of response time to host writes, but is vulnerable to partial data loss in the case of failure.